Fitting GIUH Probability Distributions to Observed Data Using Scalar Values to Facilitate Hydrologic Model Calibration
Abstract
A Geomorphic Instantaneous Unit Hydrograph (GIUH) is widely used in hydrologic models including the Conceptual Functional Equivalent (CFE) of the National Water Model currently under development by the Office of Water Prediction (OWP). Various methods exist for computing GIUH for a given basin including from stream order or from Digital Elevation Model (DEM) data, however no model will take all factors influencing basin response into account-for example, soil characteristics, vegetation, land use and urbanization can substantially affect response curves in ways that are difficult to model, and foreknowledge of such parameters is often difficult to obtain. This leads to a need to calibrate GIUH-dependent models to better fit forecast basin response to observations. This becomes especially important when applied to many small catchments where response timings may consist of a very small curve of coarse ordinates (the National Water Model will use a network of catchments of 3 to 15 sqkm in size).
Further, GIUH derivation methods may be numerical in nature leading to discrete ordinates as output, or the probability distribution information may be captured as discrete ordinates at a coarse time resolution (e.g. hourly) and thus lose the analytically derived function underlying the distribution. These factors lead to a difficulty in calibrating GIUH-dependent models, hindering the ability to tune the model and improve performance. We examine methods of resampling discretized probability distributions as applied to use in calibrating hydrologic models composed in whole or in part by GIUH models. Because the ordinates must sum to 1 and the number of ordinates for a given basin may be arbitrary, making direct calibration of the values difficult, we limit the tunable parameters to a small, fixed number of scalar values (as few as one) that are used to deform and resample a baseline GIUH, facilitating calibration to observed basin response in concert with the calibration of other, similar scalar model parameters and with existing multidimensional search algorithms such as Gauss-Marquard-Levenberg (GML), Dynamically Dimensioned Search (DDS), or Shuffled Complex Evolution (SCE). We demonstrate and quantify the effect of the calibrated models on basin response metrics such as peak timing as compared to observed response.- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFM.H45I1496W